This paper presents an Automatic Speech Recognition (ASR) system, in the Gujarati language, developed for Low Resource Speech Recognition Challenge for Indian Languages in INTER-SPEECH 2018. For front-end, Amplitude Modulation (AM) features are extracted using the standard and data-driven auditory filterbanks. Recurrent Neural Network Language Models (RNNLM) are used for this task. There is a relative improvement of 36.18 % and 40.95 % in perplexity on the test and blind test sets, respectively, compared to 3-gram LM. Time-Delay Neural Network (TDNN) and TDNN-Long Short-Term Memory (LSTM) models are employed for acoustic modeling. The statistical significance of proposed approaches is justified using a bootstrap-based % Probability of Improvement (POI) measure. RNNLM rescoring with 3-gram LM gave an absolute reduction of 0.69-1.29 % in Word Error Rate (WER) for various feature sets. AM features extracted using the gammatone filterbank (AM-GTFB) performed well on the blind test set compared to the FBANK baseline (POI>70 %). The combination of ASR systems further increased the performance with an absolute reduction of 1.89 and 2.24 % in WER for test and blind test sets, respectively (100 % POI).
This paper presents our work on end-to-end (E2E) system development for multiple array track in the CHiME-5 challenge 2018. In particular, we propose to use E2E Lattice Free Maximum Mutual Information (LF-MMI) for acoustic modeling. For front-end, Mel Frequency Spectral Coefficients (MFSC) and Power Normalized Spectral Coefficients (PNSC) features are used. We employ delay-and-sum beamformer for speech enhancement of training and development data. The Recurrent Neural Network Language Model (RNNLM) rescoring is also explored along with 3-gram language model. Our E2E LF-MMI Time Delay Neural Network (TDNN) system performed better than the E2E system provided in the challenge with an absolute reduction of 10.95 % in WER. The final system combination further reduces the WER to 78.63 %. Hence, our proposed system combination captures complementary information due to various E2E systems trained on full training data, beamformed data and using MFSC and PNSC features, respectively.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.